Capturing a snapshot of user stacks

[Wim Lewis]

I’m looking for a tool which can capture all (or a selected subset) of the user stacks on the system, even for tasks that are blocked in a syscall at the time.

Dtrace can capture a user stack but only when an event happens in that task; and procstat -k can capture all the stacks but only the kernel stacks. Is there a way to, in effect, have a dtrace probe fire in every blocked thread? Or another way to get this information?

A little explanation: I’m trying to discover why a system occasionally becomes slow. When this happens, I want to capture some profiling information for later perusal. I can get a lot of information from dtrace (a la http://www.brendangregg.com/FlameGraphs/offcpuflamegraphs.html ) and pmc, but that won’t tell me anything about a process that’s blocked the whole time, or is blocked at the start of the interval but unblocks partway through (as is likely to be the case if I only start collecting data once the problem occurs).

Any ideas?


Wim Lewis / wiml at omnigroup.com